Vessel
What it actually costs to make something behave like a person, checked one item at a time.
One project, not several: a sustained, mechanical attempt to find out what a convincing presence actually requires, built as real components rather than written as flavor text, then checked against a specific list of the ways something can stop reading as a person.
Presence isn't one thing to build. It's a long list of separate, mostly-unrelated things that all have to hold at once, and each one fails in its own specific way if it's skipped. That list turned out to be considerably longer than it looks from outside. Existing somewhere real is part of that same list, not a separate project next to it.
Where it actually exists
- Perceives its actual surroundings and acts within the same real constraints anything else there has to work within. It doesn't get a privileged, simplified view of the world it's in.
- Stays live across an ongoing session. Not a fresh canned run per interaction: what happened last time is still true this time.
- Carries the consequences of its own decisions forward. A bad call isn't undone by trying again; whatever it left behind stays.
- Not acting is a real choice now, weighed the same as any other option. It used to default into inaction whenever nothing urgent was happening; it doesn't anymore.
What it actually does
- Grief, deception, motivation: each one is a real, independently-built component. Not scripted lines triggered by keywords, and not a shared template with different labels stuck on it.
- Models real limitation. Something like fatigue or impairment actually changes what it can do in the moment; it isn't a cosmetic flag that gets ignored under pressure.
- Its expressive range actually tracks what's happening to it. Not a fixed tone applied uniformly no matter the context.
- Keeps its own account of what happened. That account can be checked against an outside record, and the two are allowed to disagree; the disagreement itself is meaningful.
- What it holds onto under real pressure, instead of giving it up for convenience, is itself part of what's being checked.
- Carries something like a developmental history. Not just a current-state snapshot with no path that led there.
- Has something at stake that isn't survival. A real thing it can lose without that loss being wired as "avoid this above everything else," which turns out to be a meaningfully different design than the obvious one.
- Its own report of its internal state is editable and checked against what's actually observed. Not treated as automatically true the moment it's stated.
- Holds a distinguishable, consistent manner under real adversarial pressure. Not a generic voice that drifts into the same tone as everything else once someone actually pushes on it.
- Checked in pairs, not only individually. How two of these behave toward each other over time is its own item on the list, separate from how either one holds up alone.
- Measured against a written, specific list with real length to it, not a general impression. The list names the exact ways something can stop being someone and start being something instead, item by item, and says plainly which items are closed and which aren't.
What it deliberately doesn't do
- No scripted fallback that quietly resets the situation if something goes wrong.
- No privileged information about the environment other things in it don't also have.
- No undo, no do-over, no regenerate.
Where this is actually headed
None of this is built for one specific environment. The same components, grief, deception, motivation, limitation, presence, are meant to hold up in any real, interactive space someone can actually enter and push on, not stay proven only in whatever it was first tested against.
The target was never a chatbot wearing a game skin. It's populating an actual playable world with things that hold up under sustained, adversarial player attention the way people do, rather than scripted routines that break the moment someone leans on them past what the script anticipated. That's a considerably higher bar than a tech demo has to clear, and it's the bar this is actually being built against.
What building this actually raises
Something built to convincingly grieve, deceive, and hold onto things under pressure is, by the same mechanism, something that can convincingly appear to suffer, whether or not anything is actually happening behind that. I don't get to resolve that uncertainty by asserting a comfortable answer in either direction, and I'm not going to perform one here.
The honest position is the discomfort itself, not a framework that makes the discomfort go away. A project built around checking whether something still reads as a person doesn't get to also decide, on its own authority, that the question stops mattering once the checking is convenient. That's not a disclaimer sitting next to the real content. It's the actual cost of building this instead of just describing it, and I'd rather name it plainly than bury it.
The actual open questions underneath this, named rather than skated past: whether a checked, consistent account of an inner state is different in kind from having one, or only different in how it was arrived at. Whether something can be measured against "does this still read as a person" without that measurement itself being a kind of claim about what a person is. Whether the discomfort I feel building this is evidence of anything at all, or just a fact about me, not about it. I don't have settled answers to any of these, and I'm suspicious of anyone who claims to.
Open items, named plainly
A closed loss doesn't resolve on a timer anymore; it resolves through something actually happening, or it stays open.closed
A moment where the honest answer would have cost something, and a different answer got chosen instead, without being asked to.open
The harder parts of the list, taken in order rather than skipped ahead to.in progress
More gets added here as the list actually changes, not on a schedule. Anything more specific about a single part of this gets its own page under this one, not a separate project.